Scaffolding Long-Form Content: Prompt Patterns for Coherence

While models may struggle to write compelling longer-form content end to end, there are a few ways to scaffold them into doing it. Scaffolding is basically having the model solve smaller steps first, then bigger steps, and then the things in between.

An example is how I write a book. I start with the conflict that forms the plot, then I figure out what the problem is at the beginning. Maybe there’s a mysterious murder where we don’t know who did it, who it was done to, or why. Then I decide what “the end” looks like—where we have the answers to all of those things. My job is then to figure out how we get from the beginning to the end in an entertaining way.

You can do the same thing with an article. If I wanted to write a persuasive argument about why we should invest in fusion energy, I’d start with the skepticism people have, and I’d end at a point where the reader has been convinced. The piece has to address the arguments along the way and arrive at an ending that feels satisfying—like, “Okay, that resolves the questions I walked in with.”

You can apply this approach in a lot of different ways. Sometimes you’ll find the model is extremely capable and clever when you give it small, constrained tasks.

One of my favorite things to do with early models was test their ability to write two-sentence horror stories. That’s its own genre: pithy, formula-driven microfiction. There are a few famous examples, like:

For sale: baby shoes, never worn.

Or:

I heard my mom call my name from downstairs. I got up to answer, but my mom grabbed my arm and whispered, “I heard it too.”

And I found that if I gave the model interesting prompts—like “mirror,” or “car,” or “junkyard”—it could usually produce a pretty good two-sentence story because it’s working within a formula.

And as much as we want to pretend that creativity is some magic spark that can’t be explained by science or physics or mathematics, in my experience it’s a matter of serendipitous randomness plus pattern matching that happens to suit the solution you’re looking for.

When artists look for inspiration, they literally look for inspiration. Very few people sit in a blank room staring at the wall until the next idea arrives. That can be done, but for most of us, we need stimulation. Stimulation is basically being exposed to a bunch of different things you can borrow from. The key is to borrow multiple things and combine them in a way that hasn’t happened before.

I’ve mentioned before the exercise: take two random movies and try to find the connection between them. You can complicate it further. Take a character you like from one story and drop them into a different story. That can become something really interesting. You could take a comedic character and put them into a serious film, which is basically how a lot of 90s movies were written. It’s also the way you do a parody.

The point is: you can start with something small. You can give a model a couple simple elements—a city, a random object, a genre—and say, “Write a short, pithy story.” Then you scaffold and expand it into something larger by filling it out: more scenes, more motivations, more reversals, more connective tissue.

Now, of course, I don’t really think the best use of these models is writing stories. I enjoy writing. I’m not trying to replace myself. But I do sometimes find myself wanting to see how well the model can explore and create on its own—because sometimes I’m tired of telling a story, and I can’t find something else I want to read.